cv

Graduate Research Assistant — Robot Learning

RobotiXX Lab, George Mason University

Student-aware curricula over tasks, rewards, and simulation conditions; advised by Dr. Xuesu Xiao

• Developed GACL with VAE task representations, learner-performance history, and grounded task sampling for PPO in 128 parallel Isaac Gym environments. Reported simulation success: 81.85% navigation and 79.21% quadruped locomotion versus CLUTR at 76.67% and 74.65% (first author, IROS 2025)
• Co-developed Reward Training Wheels to adapt auxiliary-reward weights while retaining the primary objective. Reported off-road simulation success: 76.67% versus 34.44% with expert rewards; physical completion: 5/5 versus 2/5 trials (co-first author, IROS 2025)
• LUCID (preprint, 2026): execution-informed domain-randomization curricula for humanoid motion tracking. The study reports 88.9% versus 76.8% full-randomization simulation completion against filtered-error PI; physical G1 trials with 40 ms added delay: 38/60 versus 23/60
• Third author on Moving Through Clutter (2026), a VR data-collection and evaluation framework for scene-aware humanoid locomotion with 348 trajectories across 145 3D scenes
• Developing a C++/ROS 2 policy-inference stack for Unitree G1 with paired ONNX residual/base policies, metadata-driven observation construction, temporal history, normalization, and Isaac Lab–MuJoCo parity diagnostics
• Co-authored RL-based Adaptive Dynamics Planning (4th author, ICRA 2026) and Decremental Dynamics Planning (3rd author, IROS 2025); the DDP-based RobotiXX system placed 2nd in both the simulation qualifier and physical finals of the 2025 BARN Challenge

Software Development Engineer Intern — Amazon Aurora PostgreSQL

Amazon Web Services (AWS)

Aurora PostgreSQL query-execution performance and compatibility

• Implemented and extended Adaptive Join in C for Aurora PostgreSQL, adding support for additional join types and expanding supported query coverage
• Built join benchmarks and synthetic-data generators; used statistical comparisons to investigate regressions across engine changes

Software Development Engineer Intern — RDS Proxy

Amazon Web Services (AWS)

Statistical performance testing and visualization infrastructure

• Built a Streamlit application that unified multi-region RDS Proxy performance comparisons and regression investigation
• Implemented regression detection using Welch's t-test, power analysis, and Bonferroni correction; integrated CloudWatch metrics into reproducible performance dashboards
• Developed adaptive test selection with Thompson sampling and Bayesian optimization to prioritize informative test configurations

Research Assistant — 3D Perception and AR

Computational Engineering and Robotics Lab, Carnegie Mellon University

Deep learning for 3D augmented-reality scene completion

• Built an AR scene-inpainting pipeline using GAN image completion plus RANSAC and DBSCAN point-cloud segmentation

Research Assistant

Biorobotics Lab, Carnegie Mellon University

Computer vision for recycled-material classification

• Built a CNN and OpenCV pipeline for real-time recycled-paper classification

Ph.D. in Computer Science

George Mason University

Research focus: curriculum learning and reinforcement learning for robotics

• Advanced Machine Learning
• Deep Learning
• Reinforcement Learning
• Computer Vision

M.S.

Carnegie Mellon University

Mechanical Engineering

• GPA: 3.94/4.0
• Machine Learning
• Deep Learning
• Computer Vision
• Deep Reinforcement Learning and Control

B.S.

University of Cincinnati

Mechanical Engineering

• Magna cum laude

Adaptive Dynamics Planning for Robot Navigation

IEEE International Conference on Robotics and Automation (ICRA)

Fourth author. Reinforcement learning adapts dynamics fidelity to environmental complexity for safer, more efficient navigation.

Moving Through Clutter: Scaling Data Collection and Benchmarking for 3D Scene-Aware Humanoid Locomotion via Virtual Reality

arXiv preprint

Third author. VR data-collection and evaluation framework and benchmark for scene-aware humanoid locomotion; 348 trajectories across 145 cluttered 3D scenes.

ColorMap-VIO: A Drift-Free Visual-Inertial Odometry in a Prior Colored Point Cloud Map

IEEE Robotics and Automation Letters (RA-L)

Fifth author. Drift-bounded visual-inertial localization against a prior colored point-cloud map.

Decremental Dynamics Planning for Robot Navigation

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Third author. Dynamics fidelity decreases along the planning horizon; the resulting system placed 2nd in both phases of the 2025 BARN Challenge.

Reward Training Wheels: Adaptive Auxiliary Rewards for Robotics Reinforcement Learning

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Co-first author with Tong Xu. Adaptive auxiliary-reward weighting; reported off-road simulation success of 76.67% versus 34.44% with expert rewards, and physical completion of 5/5 versus 2/5 trials.

GACL: Grounded Adaptive Curriculum Learning with Active Task and Performance Monitoring

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

First author. Adaptive task generation grounded in reference tasks; reported navigation/locomotion simulation success of 81.85%/79.21% versus CLUTR at 76.67%/74.65%.

II-NVM: Enhancing Map Accuracy and Consistency with Normal Vector-Assisted Mapping

IEEE Robotics and Automation Letters (RA-L)

Fifth author. Normal-vector-consistent SLAM mapping resolves the double-sided mapping issue.

Humanoid Policy Inference Prototype

Ongoing C++/ROS 2 prototype for Unitree G1 motion-policy inference

• Added paired ONNX residual/base inference, metadata-driven observation assembly, history buffering, normalization, and Isaac Lab–MuJoCo parity diagnostics
• Documented validation gates explicitly; no physical-deployment claim is made

LUCID: execution-informed domain randomization

Preprint, 2026. Latent-Understanding Curriculum for Informed Domain Randomization in Humanoid Motion Tracking.

• A frozen temporal encoder summarizes command–execution discrepancy; a bounded PI scheduler with return-based backoff paces shared randomization intensity
• Study evaluates held-out motion tracking, an unseen 60 ms delay, and physical G1 trials across four motions, three checkpoints, and five repetitions per condition and method

RobotiXX Team — 2nd Place, Simulation and Real-World Phases, 2025 BARN Challenge

IEEE ICRA 2025

Official team result for the DDP-based navigation system.

Teaching Assistant — Introduction to Programming

George Mason University

Fall 2023

Teaching Assistant — Artificial Intelligence and Machine Learning

Carnegie Mellon University

Fall 2022

Teaching Assistant — System Dynamics, Fluid Dynamics and Engineering Models

University of Cincinnati

Fall 2020